1. Dialogue Modeling and Dialogue Management Frameworks Svetlana Stoyanchev Seminar on SDS, Columbia 2/16/2015

2. – Dialogue modeling: formal characterization of dialogue, evolving context, and possible/likely continuations – Theoretical approach – Dialogue management SDS module concerned with dialogue modeling – Practical implementation Seminar on Spoken Dialogue Systems, Columbia 2015 2

3. Outline Plan-based dialogue Speech Act theory Dialogue Annotations Information State Approach DM Frameworks Seminar on Spoken Dialogue Systems, Columbia 2015 3

4. Plan-based Models of Dialogue Theory originators: Cohen, Allen, Perrault This theory was developed 1979 - 1996 Dialogues are treated as a special case of rational (non-communicative) behavior – Dialogue is part of the behavior, not just a communicative interface Participants are rational agents Seminar on Spoken Dialogue Systems, Columbia 2015 4

5. Plan-based Dialogue Theory Agent’s actions are based on BDI: Belief, Desire & Intention Dialogue participants are rational agents Environment: kitchen Participants: A and B A is a robot that can move but is not familiar with the kitchen environment B can not move but knows where everything is Seminar on Spoken Dialogue Systems, Columbia 2015 5

6. Knowledge and wants Knowledge – KNOW (B, P) B knows that P is true Wants: Make Tea P1: teacup is in a cupboard P2: teapot is on a countertop P3: water is in the sink etc Participant: A Participant: B Seminar on Spoken Dialogue Systems, Columbia 2015 6

7. Make tea Participant: A Participant: B Where is a tea cup? Goals Plans: Goals Plans Infer A’s goal/plan (making tea) Infer A’s goal/plan (making tea) Personal goals Its in on the lower shelf in a cupboard above the sink Respond to A Boil water find cattle pour water turn cattle on Get a tea cup locate cup retrieve cup Get a tea bag locate retrieve Combine Boil water find cattle pour water turn cattle on Get a tea cup locate cup retrieve cup Get a tea bag locate retrieve Combine Provide additional info And the sugar is above it Plan construction & knowledge share Seminar on Spoken Dialogue Systems, Columbia 2015 7

8. Speech Acts Utterances are actions that update dialogue context A link between the mental states of agents and purposeful communication Roots in Philosophy – Austin (1962), Searle (1975) Effects of an utterance – Locutionary Words that constitute an utterance – Illocutionary Request (semantics and intent of an utterance) – Perlocutionary Underlying intention, effect, pragmatics Seminar on Spoken Dialogue Systems, Columbia 2015 8

9. Speech Acts “Are there any Flights from Boston” Effects of an utterance – Locutionary Words that constitute an utterance – Illocutionary Request information about Flights from Boston – Perlocutory Effect: requesting a clerk to provide information Seminar on Spoken Dialogue Systems, Columbia 2015 9

10. Speech Acts Are used for planning in dialogue – A dialogue plan is dynamically constructed using Agent goals Agent knowledge Speech act preconditions Speech act effects Seminar on Spoken Dialogue Systems, Columbia 2015 10

11. Common Speech Acts Inform Request Response Yes/No question Seminar on Spoken Dialogue Systems, Columbia 2015 11

12. Common Speech Acts Inform – Speaker (S), Hearer (H), Preposition (P) – Preconditions: KNOW(S,P) and WISH (S, P, H) – Effect: KNOW(H,P) – Speaker informs the hearer of something by causing hearer to believe that the speaker wants them to know something Seminar on Spoken Dialogue Systems, Columbia 2015 12

13. Why aren’t logic and plan‐based approaches used more today? Most systems are simpler, single domain, limited task, don’t need complex reasoning Focus on robustness/speed Need to deal with uncertainty in input processing Speech act theory is adapted for more structured (and limited) dialogue Seminar on Spoken Dialogue Systems, Columbia 2015 13

14. Dialogue Annotation Schemes Schemes for annotating dialogues Used for : – building statistical models of dialogue prediction of next dialogue act in a live system with a combination of features – from the current utterance – Previous utterances (dialogue history) – Observer system that analyses human dialogue Can you think of examples? Seminar on Spoken Dialogue Systems, Columbia 2015 14

15. Dialogue Annotation Schemes Schemes for annotating dialogues Used for : – building statistical models of dialogue prediction of next dialogue act in a live system with a combination of features – from the current utterance – Previous utterances (dialogue history) – Observer system that analyses human dialogue Meeting assistant Dialogue summarization Seminar on Spoken Dialogue Systems, Columbia 2015 15

16. DAMSL generic annotation schema Dialogue Markup in Several Layers Core & Allen 1997 Addresses the problem with single label of speech acts – Because of multi-purpose nature of utterances DAMSL uses 3 layers – Each utterance can have multiple labels Inform-accept Inform Speech Act Seminar on Spoken Dialogue Systems, Columbia 2015 16

17. DAMSL generic annotation schema 3-layer scheme – Forward communicative functions Traditional speech act theory – Backward communicative function How current utterance relates to the previous dialog – Accept/reject a proposal – Confirm understanding – Answer a question – Utterance Features Communicative process vs subject Seminar on Spoken Dialogue Systems, Columbia 2015 17

18. DAMSL scheme Forward Functions Backward Functions Utterance Features Seminar on Spoken Dialogue Systems, Columbia 2015 18

19. DAMSL scheme Forward Functions Backward Functions Utterance Features Seminar on Spoken Dialogue Systems, Columbia 2015 19

20. DAMSL example Seminar on Spoken Dialogue Systems, Columbia 2015 20

21. Evaluation of Annotation Scheme Do annotators agree on the tags? – Multiple annotators annotate the same dialogue and compare the agreement Kappa Statistic to measure inter-annotator agreement (IAG) – PA = % agreement – PE = % expected agreement (depends on the number of possible tags).67 is considered reliably good Seminar on Spoken Dialogue Systems, Columbia 2015 21

22. DAMSL scheme Kappa IAG Forward Functions Backward Functions Utterance Features.66.70.15.57.42.60 Seminar on Spoken Dialogue Systems, Columbia 2015 22

23. Other generic annotation schemes Carletta (1997) H. Bunt (2013) – ISO standard – Motivated by all previous annotation schemes – Multifunctional (as DAMSL) A: Henry, can you take us through these slides? H: Oh Okay, just ordering my notes – A assigns the next turn to H – A formulates indirect request – Not limited to 3 dimensions: Each utterance is labeled with – 1 general purpose function – up to 9 dimension-specific functions Seminar on Spoken Dialogue Systems, Columbia 2015 23

24. General function of DIT annotation scheme Seminar on Spoken Dialogue Systems, Columbia 2015 24

25. General function of DIT annotation scheme Seminar on Spoken Dialogue Systems, Columbia 2015 25 U: Who answered the door?

26. General function of DIT annotation scheme Seminar on Spoken Dialogue Systems, Columbia 2015 26 U: Did Tom or Mary answer the door?

27. Dimensions of DIT scheme Task Feedback Turn management Time management Discourse structuring Social obligation management Seminar on Spoken Dialogue Systems, Columbia 2015 27

28. Examples of dimension-specific annotations Seminar on Spoken Dialogue Systems, Columbia 2015 28

29. How are annotation frameworks used? Analysis of human dialogues (meetings, customer service dialogues with live agents, non-task- oriented dialogues) – Analyzing dialogue structure & building predictive models Manual annotation of DA Build automatic models to classify DA Building models for SDS applications – Annotate user responses – Build predictive models for DA Seminar on Spoken Dialogue Systems, Columbia 2015 29

30. Comments on generic annotation schemes Generic schemes are designed to handle general (human) conversation SDS scheme are simplified for application purposes Application-specific DA schemes Seminar on Spoken Dialogue Systems, Columbia 2015 30

31. Dialogue Management

32. Dialog system components Voice input Hypothesis (automatic transcription) Text Speech Language Model/Grammar Acoustic model Grammar/Models Generation templates/ rules Logical form of user ’ s input Logical form of system ’ s output Seminar on Spoken Dialogue Systems, Columbia 2015 32

33. Dialogue Manager (DM) Is a “brain” of an SDS Decides on the next system action/dialogue contribution SDS module concerned with dialogue modeling Seminar on Spoken Dialogue Systems, Columbia 2015 33

34. Dialogue Management Address how SDS is implemented – Reusability of components Approaches: – Frame-based approach (structural) – Information State approach (Traum & Larsson 2003) Seminar on Spoken Dialogue Systems, Columbia 2015 34

35. Information State Approach Formalizes theories of dialogue – Speech act Combines different theoretical approaches to SDS – Planning (more flexible and complex) – Structural (simple scripted dialogues) Leads to better engineering of SDS components – Separate development of modules – Facilitates reuse of components Seminar on Spoken Dialogue Systems, Columbia 2015 35

36. ISU Architecture Seminar on Spoken Dialogue Systems, Columbia 2015 36

37. Information State DM Formalize DM function as Information State Update – Identify relevant aspects of information How are they updated What controls this updates Dialogue context – Current state of the system including Known information Role of DM – Update dialogue context – Provide context-dependent expectations – Interface with task/domain processing – Decide what context to express next Seminar on Spoken Dialogue Systems, Columbia 2015 37

38. Information state Private: – Belief set – Agenda: stack of actions Shared: – Belief set – QUD: stack of questions – LM: move Seminar on Spoken Dialogue Systems, Columbia 2015 38

39. Declarative representation of system behavior – SDS designer writes update rules Seminar on Spoken Dialogue Systems, Columbia 2015 39

40. Discussion questions for Traum& Larsson paper Do the dialogue moves and update rules have to be tailored for each specific? The authors mention that information state- based theory of dialogue includes informational components, such as common ground, beliefs, intentions, etc. How are these concepts extracted, and then represented in this approach? Seminar on Spoken Dialogue Systems, Columbia 2015 40

41. SDS Trends 1966 Eliza – string pattern match DM 1970 – 2000: Plan-based systems, Speech act theories 2000 – 2010: Information State Update, Frame- based SDS (Communicator DARPA project) 2010 – 2014: Using web as data source, template- based DM, virtual assistants on phones (Siri, Cortana, Google Now), in-car SDS 2015 using inference engines on the back-end ( What’s next? Seminar on Spoken Dialogue Systems, Columbia 2015 41

42. Presentations Seminar on Spoken Dialogue Systems, Columbia 2015 42 **David Traum and Staffan Larsson, The Information State Approach to Dialogue Management. in Current and New Directions in Discourse and Dialogue, Ed. Jan van Kuppevelt and Ronnie Smith, Kluwer, pages 325--353, 2003. Discussion papers: Ben Hixon, Rebecca Passonneau Open Dialogue Management for Relational Databases NAACL 2013 Presenter: Jee Hyun Wang Pierre Lison. Model-based Bayesian Reinforcement Learning for Dialogue Management. In Proceedings of the 14th International Conference of the Speech Communication Association (Interspeech 2013), Lyon, France, 2013. Presenter: Edward Li

43. Find Project Partner 2/16 Find a project partner 2/23 Project descriptions due Considerations when finding partners: – What is your focus (system vs. research question) – Which topics – Goals to submit a paper Seminar on Spoken Dialogue Systems, Columbia 2015 43